AGENT SKILL

wandb

Use W&B's Python SDK and CLI for experiment tracking, offline/local workflows, artifacts and registries, Public API exports, automations, sweeps, and Launch jobs.

License 许可
MIT
最近更新
2026-09-03
Stars
★ 351
Forks
27

01 前言 Front Matter

name
wandb
license
MIT
metadata
{ "disco-role": "operating" }
description
Use W&B's Python SDK and CLI for experiment tracking, offline/local workflows, artifacts and registries, Public API exports, automations, sweeps, and Launch jobs.
disable-model-invocation
true

02 资源文件 4

03 SKILL.md 正文

W&B Repo Skill

Use this skill when a task involves the wandb Python package, the wandb/wb CLI, experiment tracking, local/offline run management, artifacts/model registry, Public API queries, automations, hyperparameter sweeps, or W&B Launch jobs.

Install and Verify

For ordinary users, install the public package:

python -m pip install wandb
python - <<'PY'
import wandb
print(wandb.__version__)
print(wandb.init)
PY

Use optional extras only when the workflow needs them:

  • Cloud artifact storage: wandb[aws], wandb[gcp], or Azure storage/identity packages.
  • Rich media logging: wandb[media] plus framework/media libraries used by the data.
  • Sweeps/Launch/cloud queues: wandb[sweeps], wandb[launch], and provider SDKs or container tooling required by the target backend.
  • Workspaces or model helpers: wandb[workspaces] or wandb[models] when a task explicitly uses those surfaces.

For a local source checkout, expect compiled helper artifacts such as wandb-core to be built. If source installation fails, read references/troubleshooting.md before broadening dependencies.

Run the bundled environment check when you need a safe local diagnostic:

python scripts/check_wandb_environment.py --check-cli

Route by Workflow

  • Experiment tracking in Python: Use sub-skills/experiment-tracking/SKILL.md for wandb.init(), Run.log(), config, summary, tables/media/plots, metric axes, notebooks, offline/disabled modes, and tracking smoke checks.
  • CLI and local/offline workflows: Use sub-skills/cli-and-local-workflows/SKILL.md for wandb login, wandb init, wandb status, wandb offline/online/disabled, wandb sync, local run directories, environment variables, cache cleanup, Docker wrapping, and self-hosted host configuration.
  • Artifacts and registries: Use sub-skills/artifacts-and-registries/SKILL.md for wandb.Artifact, Run.log_artifact(), Run.use_artifact(), artifact aliases/versions, downloads, wandb artifact ..., storage references, registry linking, and model registry troubleshooting.
  • Public API and automations: Use sub-skills/public-api-and-automation/SKILL.md for wandb.Api, querying runs/projects/sweeps/files/reports/artifacts, exporting history, pagination, and automation events/filters/actions/scopes.
  • Sweeps and Launch: Use sub-skills/sweeps-and-launch/SKILL.md for sweep YAML/JSON, wandb.sweep(), wandb.agent(), wandb sweep, wandb agent, Launch jobs, queues, launch agents, resource backends, and structural config validation.

Shared References

  • references/repo-provenance.md: read before deciding whether this skill is current for a different W&B checkout or before running a refresh.
  • references/troubleshooting.md: use for cross-cutting install/import, authentication, base URL, optional dependency, source-build, and local service issues.
  • scripts/check_wandb_environment.py: safe diagnostics for installed package version, importability, CLI availability, optional extras, and selected environment variables with secrets redacted.

Safe Defaults

  • Prefer with wandb.init(...) as run: so runs finish even when code raises.
  • Use mode="offline" for credentials-free testing and local-only capture; sync later with CLI guidance.
  • Never print API keys or .netrc contents. Treat WANDB_API_KEY as present/absent only.
  • Use full W&B paths such as entity/project/run_id or entity/project/artifact:alias when defaults could resolve to the wrong account.
  • For destructive commands such as wandb sync --clean or cache cleanup, explain what will be deleted and ask before running them.
  • Do not invent cloud credentials, queues, Kubernetes contexts, container registries, IAM roles, bucket names, or self-hosted server URLs.

Common Cross-Workflow Paths

  • If a training script needs both tracking and datasets/models, start with experiment-tracking, then route artifact inputs/outputs to artifacts-and-registries.
  • If an offline run must be uploaded later, instrument with experiment-tracking and use cli-and-local-workflows for wandb sync.
  • If a report/export task references existing runs, use public-api-and-automation; if it needs to create future runs, route back to experiment-tracking.
  • If a sweep trial logs metrics or artifacts, use sweeps-and-launch for orchestration, experiment-tracking for trial logging, and artifacts-and-registries for data/model lineage.

Freshness Check

This skill was generated from repository evidence and live package inspection. Read references/repo-provenance.md if the current checkout, package version, public CLI/API, tests, or docs have changed. If the provenance differs materially, refresh the skill from the repository before relying on exact signatures or command behavior.

04 同仓库更多 Skills

VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-triton-opt

Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-solve

END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-skill

PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-pattern-fusion

Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-orchestrate

Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-feedback

Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).

★ 351